What Do They Think? Undergraduate Perceptions of Student-AI Use

Synthesis: Niebuhr, Gómez-Franco, Pickus, Finch and Badilla-Quintana report an exploratory survey of 214 undergraduates at two public North American universities, one small Midwestern and one large Western institution, on how they read their peers' use of generative AI. Two questions organize the study: whether students think AI encourages academic dishonesty, and whether they see it as conferring an unfair advantage. A majority — 71.6% — said it encourages dishonesty, while 56.6% said it does not confer an unfair advantage, and 86.6% believed most of their peers use AI for university work. Logistic regression tied the dishonesty judgment to clear university guidelines, perceived career value and in-person study, while the open-ended answers locate the real cost in learning rather than in grades.

Key Findings

1. Most students read AI as a dishonesty risk. 141 respondents (71.6%) said AI encourages academic dishonesty, against 56 (28.4%) who said it does not.
2. Unfair advantage divided the sample. 86 respondents (43.4%) saw an unfair advantage for users, while 112 (56.6%) did not.
3. Peer use was treated as the norm. 174 respondents (86.6%) believed most of their peers use AI for university work, and students who used AI themselves were about three times as likely to say so (odds ratio = 3.02).
4. Clear guidelines went with less concern. Students who believed their university had developed clear AI guidelines had lower odds of saying AI encourages dishonesty (b = –0.61, odds ratio = 0.54).
5. Career value and modality both mattered. Believing AI skills would help their career lowered the odds of a dishonesty judgment (odds ratio = 0.39), while students taught mostly in person were 2.5 times more likely than online students to say AI encourages dishonesty (odds ratio = 2.49).
6. Two-thirds had used the tools. 66% reported using AI for school-related work, most commonly for small assignments such as discussion boards (64% of users), while 34% reported no use at all.

A majority reads AI as a risk to integrity

The headline is unambiguous: 71.6% of respondents said using AI encourages academic dishonesty. The open-ended explanations show what sits behind the number. "Cheating", "plagiarism", missing "attribution" and "laziness" recurred, as did the claim that AI lets students "get away without learning the actual material" and "takes away from the cognitive demand". Some students allowed that the tool helps with brainstorming while describing the temptation to use it for everything as a "slippery slope".

The minority who answered no placed the fault in the user rather than the technology. Keywords there were "user", AI as a "tool", and equal "access"; several noted that citing or crediting AI is the relevant question. The distinction matters for policy: the two groups disagree less about what AI does than about where responsibility for misuse sits, and a rule that targets the tool will read differently to each.

Notably, the paper reports 141 respondents choosing yes and 56 choosing no, but 130 and 51 of those respectively wrote an explanation, so the qualitative material covers a large majority of each camp rather than a self-selected handful.

Advantage: a split verdict about the short and long run

On unfair advantage the sample divided 86 to 112, and the reasoning on both sides converged on the same distinction between short-term grades and long-term capability. Students who saw an advantage pointed to marks: "they will likely have higher assignment grades, but it will not give advantage when entering the workforce". One described pressure to use AI on binary-answer assignments, since abstaining "can feel like intentionally putting yourself at a potential disadvantage". Another confessed guilt at having relied on it: "I have not put in the effort as others have".

Those who saw no advantage usually argued that users harm themselves. Responses described lower-quality work, failure to "retain or learn", and risk of expulsion, and several noted that everyone has access, so the playing field is level. The authors observe that this convergence is what makes the question worth pursuing: the disagreement is not about whether AI changes outcomes but about which outcomes count.

Guidelines, modality and career outlook

Three variables were statistically significant in the dishonesty model, which explained 9% of variance. Students who believed their university had clear guidelines were about half as likely to say AI encourages dishonesty (odds ratio = 0.54). Students who thought learning AI would help their career had lower odds again (odds ratio = 0.39). Students taught mainly in person were 2.5 times more likely than online students to see AI as encouraging dishonesty (odds ratio = 2.49).

The peer-use model worked differently. Using AI oneself was the strongest correlate of believing peers do too (odds ratio = 3.02), while in-person students were less than half as likely as online students to think most peers use AI (odds ratio = 0.38). The unfair-advantage model produced no significant predictors and explained 8% of variance.

The guideline finding is the one with the clearest institutional handle, but the survey cannot say which way the arrow runs: students in universities with visible rules may simply be less suspicious, or less suspicious students may be more likely to notice rules. The authors treat the association as a signal for policy rather than as an effect estimate.

What the open-ended answers add

Two further results frame the discussion. Guidelines were perceived as split almost evenly, with 100 respondents (51%) saying universities were creating clear rules and 96 (49%) saying they were not — a near coin-flip that suggests communication is uneven across programs. And 34% of respondents reported never using AI for schoolwork, a group whose reasons were often principled rather than practical: "doesn't feel very ethical", "atrophying our brains", and 12 respondents (5.6%) citing environmental costs.

The authors close on a legitimacy argument. If enough students operate outside the rules, the rules themselves lose authority, and a cohort that believes most peers are using AI while a majority believes that use is dishonest is precisely the condition in which norms erode. That is why they argue faculty and administrators should treat student perception as evidence about the institution, not merely as an attitude to be corrected.

What this means for practice

• Instructors. State your AI rules in the syllabus and in assignment briefs rather than relying on institutional policy: students who believed their university had clear guidelines were about half as likely to say AI encourages dishonesty (odds ratio = 0.54), and only 51% of respondents thought such guidelines existed.
• Administrators. Audit how AI rules reach students across programs, because perception split 51% to 49% at two institutions in the same system, and publish the reasoning behind permitted and prohibited uses so that the "user not the tool" camp and the "tool" camp can see which rule follows from which premise.
• Assessment designers. Address the long-run argument students already make: with 56.6% saying AI confers no unfair advantage because users learn less, assessment that rewards retained capability rather than polished submission aligns with a belief students already hold.

Limitations

• The survey is exploratory and deliberately left terms such as "dishonesty" and "unfair" undefined, so respondents interpreted them differently, especially regarding short-term grades versus long-term learning.
• The sample is 214 undergraduates at two public universities in two United States regions, with 65% from the large institution and 31% majoring in education, so the distribution reflects enrollment patterns rather than a representative national sample.
• All measures are self-reported and anonymous, with no check on actual AI use; the authors note that answers were not mandatory and that item totals vary.
• The design is cross-sectional, so no causal direction can be established for the guideline, career and modality associations.
• The analysis fits separate logistic regressions with modest explained variance (8% in the unfair-advantage model and 9% in the dishonesty model), so unmeasured factors account for most of the variation.

Citation

Niebuhr, R., Gómez-Franco, L. E., Pickus, D., Finch, H., & Badilla-Quintana, M. G. (2026). What Do They Think? Undergraduate Perceptions of Student-AI Use. Open Praxis, 18(4), 702–719.